GPT-5.6 Sol Review Roundup: What Codex Is Actually Great At

GPT-5.6 Sol does not win every creative comparison. It does something more useful for many builders: it repeatedly turns a defined job into a finished artifact quickly, with fewer tokens and less ceremony.

In this review-of-reviews, Andrew Warner and Bryan McAnulty examine seven creator tests across browser games, dashboards, video editing, browser automation, writing, image prompting, skill maintenance, and a Vision Pro drum kit. The shared pattern is not "Sol replaces Fable." It is that Sol is becoming a very strong execution model inside Codex.

Main episode and commentary credit: The Next New Thing, Andrew Warner, and Bryan McAnulty. The underlying tests remain the work and opinions of Nate Herk, Claire Vo and How I AI, Peter Yang, Every, Matthew Berman, and Bijan Bowen. Each featured video is linked and embedded below.

Source Note

This article separates three evidence levels. Official facts such as model names, API prices, and OpenAI capability claims come from OpenAI and Anthropic documentation. Observed results come from the linked creators' recorded demos. Interpretation includes Andrew and Bryan's reaction, the creators' model preferences, and the JQ AI SYSTEMS routing advice below.

Link Map

The Official Baseline

OpenAI released GPT-5.6 as a three-model family: Sol for frontier work, Terra for balanced everyday work, and Luna for lower-cost volume. OpenAI positions Sol around software engineering, computer use, knowledge work, science, cybersecurity, design judgment, and long-running agentic tasks. It also offers higher reasoning settings and an Ultra mode that can coordinate parallel work.

OpenAI's benchmark charts are vendor-reported results. The creator tests matter because they ask a different question: what happens after a person gives the model a messy file, a browser, a design brief, a video, or a native-app project and then asks for a usable outcome?

The Seven Reviews, Mapped by Workload

1. Fable as Manager, Sol as Worker

Nate Herk's open-world bike-game comparison produced the clearest framing. Fable made the more imaginative, enjoyable game. Sol produced a capable result at a much lower estimated cost. Andrew and Bryan described the practical split as Fable being the manager or creative co-founder, while Sol is the worker who implements, ships, and keeps moving.

In Nate's recorded run, Fable took about 21 minutes and Sol about 23 minutes. His reconstructed task estimates were roughly $14.22 for Fable and $4.50 for Sol, with Sol using far fewer output tokens. Those session figures are directional, not invoices, but the official list prices support the broader cost gap.

The routing lesson is more valuable than the winner: use a high-judgment model to define the product, acceptance criteria, and creative bar; use Sol to execute the plan, run tools, inspect the result, and close the gap.

2. Functional Design, Not Just a Pretty Screenshot

Claire Vo's How I AI review used a custom benchmark across PRDs, prototypes, wireframes, debugging, and agentic voice. Her score weighted human taste more heavily than the automated judge. Sol was her preferred model for full-fidelity prototypes because the interfaces were not only visually differentiated; expected controls worked.

The practical signal is that Sol's design improvement appears strongest when Codex can inspect the rendered page and iterate. A screenshot can hide broken buttons, missing states, weak hierarchy, or an unusable mobile layout. Browser-backed loops force the model to confront those problems.

Claire also preferred Terra for direct PRD writing and Sonnet 5 for agentic voice. That is a healthy result: the smaller model can be the better choice when the workload rewards brevity or conversational tone instead of maximum implementation depth.

3. Video Processing and Browser Automation

Two of the strongest Sol use cases were not conventional coding. Claire dropped a video into Codex, asked it to find and cut useful moments, then refined orientation, pace, and length through follow-up prompts. The model handled the mechanical media work; she kept editorial control.

Her browser-control example was even more operational: Codex worked through a large LinkedIn inbox using explicit relevance criteria. That demonstrates why browser agents can matter, but it also raises the permission bar. Reading, filtering, and drafting are safer defaults than sending hundreds of messages without review.

Peter Yang's six-use-case comparison reinforced the same point. His model tests covered an interactive travel site, a 3D game, browser-assisted video publishing, a mobile feature, advice, and personal AI OS cleanup. The best model changed with the workload, and human review remained part of the loop.

4. Cleaning Skills and Letting Models Review Each Other

Peter's workflow-cleanup test is less flashy than a 3D game and more relevant to daily work. Mature AI systems accumulate duplicate instructions, stale skills, contradictory routing rules, and scripts nobody trusts. Sol's ability to inspect a repository, propose consolidation, implement changes, and run checks makes it useful as a maintenance model.

Another practical pattern was cross-model review. Let Sol inspect Fable's plan or output, and let Fable critique Sol's implementation. This is not automatically better: two models can agree on the same bad assumption. It becomes useful when each reviewer has a distinct job and objective checks.

# CROSS-MODEL REVIEW CONTRACT

PLANNER:
- define the user outcome
- list constraints and risks
- write acceptance tests

IMPLEMENTER:
- build against the plan
- run tests and browser checks
- report unresolved failures

REVIEWER:
- inspect the actual artifact
- challenge assumptions
- cite failed acceptance tests
- do not rewrite unless asked

HUMAN:
- approve external actions
- judge taste and business fit
- accept or reject the result

5. Strong Writing, Different Visual Taste

Every's month-long review described Sol as a strong daily collaborator for writing and knowledge work. The team used it across inbox triage, meeting context, Slack decisions, campaigns, and research. The model's value came from continuity inside the work surface, not one isolated answer.

The notable exception was image prompting. In Andrew and Bryan's discussion of the Every test, Fable produced the stronger prompt for the same image model. This is a useful reminder that the text model directing a generator changes the visual result even when the downstream image model stays constant.

Use Sol for high-volume drafting, synthesis, and implementation. Keep a visual director, taste file, reference library, or second model in the loop when the work depends on art direction rather than correctness alone.

6. Long-Running Media and Coding Work

In Nate Herk's second video, Sol researched the GPT-5.6 launch, drafted a script in Nate's voice, generated audio through ElevenLabs, created an avatar in HeyGen, edited with HyperFrames, and reviewed frames. The result shows real orchestration across specialized tools.

It also shows the budget risk. Nate reports that Ultra coordinated multiple agents and pushed the run into hundreds of millions of tokens, with a cost around $300 for a short video. The exact figure belongs to his run, but the lesson generalizes: parallel agents can save human time while multiplying machine spend.

Matthew Berman's launch review adds the model-family and autonomous-build context. His examples emphasize that Sol can sustain larger software tasks, while Terra and Luna give builders cheaper routing options for work that does not need the flagship model.

7. Native and Spatial Apps Need Feedback

Bijan Bowen asked GPT-5.6 to build a Vision Pro drum kit. The first result was a flatter 2D interpretation. After Bijan clarified that he wanted a spatial instrument, the model produced a more convincing 3D mixed-reality direction.

This is one of the most honest tests in the roundup because it includes the miss. The model was capable of the better result, but the first prompt did not force the intended medium. The improvement came from inspection and specific feedback, not from restarting with a more dramatic one-shot prompt.

Cost and Model Routing

Official API list prices checked July 12, 2026 place GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens. Anthropic lists Fable 5 at $10 input and $50 output. Sol is cheaper per token, but per-token pricing is not the final bill.

The relevant metric is cost per accepted result. A cheaper model that needs five repair cycles can cost more than an expensive model that finishes once. A powerful model that overproduces millions of tokens can also be wasteful when the task needed a simple script.

The Bigger Lesson: Iteration Beats the One-Shot Demo

The most useful examples in the roundup include a feedback loop. Claire tightened video clips. Peter compared practical workflows instead of one benchmark. Bijan corrected the spatial direction. Nate inspected cost after a highly autonomous media run. These are closer to real production than asking two models for one artifact and choosing the prettier screenshot.

1. Define: write the outcome, constraints, reference material, and acceptance tests.
2. Build: let the model create a complete first pass without constant interruption.
3. Inspect: run the app, watch the video, open the files, and check logs.
4. Feedback: describe observable failures, not vague disappointment.
5. Verify: rerun tests and confirm the fix did not create a regression.
6. Stop: end when acceptance criteria pass or the budget is exhausted.

JQ AI SYSTEMS Builder Eval Checklist

# REAL-WORK MODEL TEST

WORKFLOW:
One task we already perform every week.

INPUTS:
- same source files
- same reference examples
- same tool permissions

LIMITS:
- fixed time budget
- fixed spend budget
- no unreviewed external actions

ACCEPTANCE:
- required user flow works
- facts and citations check out
- output matches the brand or design system
- no console, server, or export errors
- mobile and desktop pass where relevant

MEASURE:
- wall-clock time
- model and tool cost
- number of repair prompts
- human review minutes
- accepted or rejected

DECISION:
Route by lowest reliable cost per accepted result.

Practical Verdict

GPT-5.6 Sol looks excellent inside Codex when the job is concrete, tool-heavy, and verifiable. It is especially compelling for implementation, functional UI, browser work, video processing, workflow cleanup, and long-running tasks where token efficiency makes repeated use affordable.

Fable 5 still earns a place when the work depends on creative leadership, higher-level judgment, image-prompt design, or difficult planning. Terra, Luna, and other cheaper models should carry routine volume. Human review remains the final layer for taste, external actions, and production acceptance.

Sources

• The Next New Thing: What Codex 5.6 is amazing at!
• The Next New Thing, Andrew Warner on X, and Bryan McAnulty on X
• Nate Herk: I Tested GPT 5.6 Sol vs Fable 5
• How I AI: GPT-5.6 Sol - Better and Cheaper Than Fable and written workflow notes
• Peter Yang: GPT-5.6 vs Claude Fable 5 Across Six Use Cases and written comparison
• Nate Herk: GPT 5.6 Sol Made This Entire Video
• Every: I Tested GPT-5.6 Sol for a Month and Every's written review
• Matthew Berman: GPT-5.6 is Finally Here
• Bijan Bowen: GPT-5.6 Is Here
• OpenAI: GPT-5.6 and GPT-5.6 Sol API model page
• Anthropic: Claude Fable 5 and Mythos 5 and Claude Platform pricing